Intel unveils brain-based neural system for AI workloads

Intel is launching a Halo point onto the world: it's a prototype of a so-called nervous system, which works more like a human brain. Intel hopes that Hala Point technology will pave the way for effective and affordable artificial intelligence.

Intel showcases Hala Point: a neural computing system based on the Loihi2 processor. Hala Point, like its predecessor Pohoiki Springs, should function like a human brain and therefore has its own unique architecture. The system is built around (digital) neurons and Hala Point has ten times as many as Pohoiki Springs. Memory and computing power are intertwined in the solution, just like the brain.

Hala Point is a high-volume server with 1,152 Loihi 2 processors crammed into a 6U chassis. The system contains 1.15 billion neurons and 128 billion synapses, distributed among 140,544 neural processor cores. The system uses 2.6 kilowatts of power and has 2304 x 86 cores to support traditional computing.

The solution provides 15 TOPS/W and is ideally suited to support deep learning networks. According to Intel, the neural approach is more powerful and efficient than traditional GPU-based AI hardware. Intel hopes that Hala Point will make training large models like GPT-4 more feasible.

Challenger for Nvidia

The solution is a direct competitor to the ecosystem created by Nvidia and appears to have an advantage on paper. However, in practice, Nvidia can rely on a wide range of software solutions. The question is to what extent AI developers are inclined to adopt a completely new architecture.

Read also

Intel unveils Gaudi 3: a direct competitor to the Nvidia H100

The Hala Point prototype is in use at the US Sandia National Laboratories and is used there to solve computational problems related to various fields of physics. The computer system is very good at solving optimization problems compared to classical computer chips, and has advantages that are also attributed to quantum computers (which do not yet exist).